HN Explorer

Show HN: Transcribe YouTube Videos

1 points by llimllib ↗ HN
I often want to read the contents of YouTube videos but I don’t want to watch them, and YouTube’s captioning leaves a lot to be desired.

So I wrote a shell script that takes a URL, downloads the video, transcribes it with whisper, and turns it into a clean HTML page for reading.

https://github.com/llimllib/yt-transcribe

11 comments

[ 3.2 ms ] story [ 37.8 ms ] thread
Leftium
Perhaps skip the whisper step if the video already has decent manual (not auto-generated) captions, like this video: https://youtu.be/i-BkN3rTK0Q

---

Another use case is being able to quickly jump to a specific spot inside a video. Could you add timestamps with links that jump directly to that point in the video?

I accomplished something similar by modifying oTranscribe:

- https://otranscribe.netlify.app/?vsl=definedefine

- https://otranscribe.netlify.app/?vsl=letter

---

Finally, I'm a windows user so a whisper.cpp version would be nice~

llimllib
good idea! I've no idea how to distinguish between auto-generated and manual captions, but I definitely should take them if available.

A timestamp flag is also a good idea.

Will noodle on a whisper.cpp version!

A question: on mac I can `brew install whisper-cpp`. Is there any equivalent way to install it on a windows machine? I haven't used windows in a very long time.

Leftium
yt-dlp parameters distinguish between auto-generated and manual YouTube captions:

    # Downloads auto-generated captions
    yt-dlp --skip-download --write-auto-sub https://youtu.be/i-BkN3rTK0Q

    # Downloads manual captions
    yt-dlp --skip-download --write-sub https://youtu.be/i-BkN3rTK0Q

    # Fails with error: "There are no subtitles for the requested languages"
    yt-dlp --skip-download --write-sub https://youtu.be/ZMklf0vUl18
Docs: https://github.com/ytdl-org/youtube-dl?tab=readme-ov-file#su...
Leftium
On Windows, I use scoop.sh: https://scoop.sh/#/apps?q=whisper

I was able to do this:

    scoop install main/whisper-cpp
    
    mkdir models
    
    ## Download model file ggml-base.en.bin to models directory above
    
    yt-dlp.exe -x --audio-format wav --audio-quality 16K -o "out.wav" ZMklf0vUl18
    
    # Wrangle wav into 16kHz format (param above did not seem to work...)
    ffmpeg -i out.wav -ar 16000 out-16kHz.wav
    
    whisper.exe out-16kHz.wav
llimllib
Presumably that used whisper's bundled tiny model, which is no better than youtube CC. A beef I have with whisper-cpp is that they totally refuse to handle model management.

With mlx_whisper, I just have to tell it to use a model and it will download it if it's not already present: https://github.com/llimllib/yt-transcribe/blob/244841f83d833...

so if I add whisper.cpp as a dependency, I also have to add huggingface-cli or something similar. It also seems like huggingface-cli is not available on scoop

Leftium
Yes, the model must be downloaded separately (see my edited comment with bash commands/comments).

The model is specified via whisper.exe `--model FNAME` parameter. By default, it looks for `models/ggml-base.en.bin`, but even that model must be downloaded separately.

So you could do this:

    # Assumes ggml-large-v3.bin model file[1] was already downloaded to models/ folder
    whisper.exe --model models/ggml-large-v3.bin out-16kHz.wav
[1]: https://huggingface.co/ggerganov/whisper.cpp/blob/main/ggml-...
Leftium
Not as convenient, but you could also have the user manually install the model, like whisper does.

Just forward the error message output by whisper, or even make a more user-friendly error message with instructions on how/where to download the models.

Whisper does provide a simple bash script to download models: https://github.com/ggerganov/whisper.cpp/blob/master/models/...

(As a Windows user, I can run bash scripts via Git Bash for Windows[1])

[1]: https://git-scm.com/download/win

llimllib
thanks for all the help, I appreciate it.
Leftium
Well, thanks to you I found out whisper generates decent audio transcriptions using a local LLM (relatively) easily, even on my 6+ year-old laptop.

(I used to upload videos to YouTube just to get the auto captions.)

I did some investigation, and it would not be difficult to convert the whisper LRC subtitle output into the format my fork of oTranscribe expects.

I already made a simple tool to convert YouTube TTML/SBV subtitle output: https://github.com/Leftium/otrgen

llimllib
that's great! whisper is awesome software.

I'm working on a golang version that links to whisper.cpp directly to maybe make porting easier/possible